focal_loss.py 文件源码

def forward(self, is_train, req, in_data, out_data, aux):

        cls_score = in_data[0].asnumpy()
        labels = in_data[1].asnumpy()
        self._labels = labels

        pro_ = np.exp(cls_score - cls_score.max(axis=1).reshape((cls_score.shape[0], 1)))
        pro_ /= pro_.sum(axis=1).reshape((cls_score.shape[0], 1))
      #  pro_ = mx.nd.SoftmaxActivation(cls_score) + 1e-14
       # pro_ = pro_.asnumpy()   
        self.pro_ = pro_
        # restore pt for backward

        self._pt = pro_[np.arange(pro_.shape[0],dtype = 'int'), labels.astype('int')]

        ### note!!!!!!!!!!!!!!!!
        # focal loss value is not used in this place we should forward the cls_pro in this layer, the focal vale should be calculated in metric.py
        # the method is in readme
        #  focal loss (batch_size,num_class)
        loss_ = -1 * np.power(1 - pro_, self._gamma) * np.log(pro_)
        print "---------------" 
        print 'pro:',pro_[1],labels[1]
        self.assign(out_data[0],req[0],mx.nd.array(pro_))